Mechanical Systems and Signal Processing 64-65 (2015) 233–244

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

A variable-coefficient harmonic balance method

for the prediction of quasi-periodic response
in nonlinear systems
B. Zhou a,n, F. Thouverez b, D. Lenoir b
a
College of Energy and Power Engineering, Nanjing University of Aeronautics and Astronautics, 29 Yudao St., Nanjing 210016, China
b
Laboratoire de Tribologie et Dynamique des Systems, Ecole Centrale de Lyon, 36 avenue Guy de Collongue, 69134 Ecully Cedex, France

a r t i c l e i n f o abstract

Article history: Quasi-periodic responses arise from various nonlinear dynamic systems under a single-
Received 1 September 2014 frequency excitation. A variable-coefficient harmonic balance method is proposed for the
Received in revised form prediction of quasi-periodic responses. The key point of this method is that the quasi-periodic
27 March 2015
response is described as a truncated trigonometric series with time-periodic Fourier
Accepted 16 April 2015
coefficients. In other words, quasi-periodic responses are treated in a “cascade” of frequency
base. Harmonic terms in the nonlinear system are separated and balanced with respect to
Keywords: each basic frequency. Numerical examples reveal that this method is efficient in predicting
Quasi-periodic response such quasi-periodic responses, which contain an unknown frequency component.
Harmonic balancing
& 2015 Elsevier Ltd. All rights reserved.
Time-periodic Fourier coefficients

1. Introduction

Coupled oscillators occur in many applications ranging from aerodynamics and structural dynamics. The existence of
steady, quasi-periodic responses arising from coupled oscillators has been observed in diverse engineering systems [1–4]. In
general, a quasi-periodic oscillation that is associated with p different internal frequencies takes place on a p-dimensional
invariant torus. Any quasi-periodic response x(t) can be expressed as a function xðtÞ ¼ xðωtÞ. The tuple ω ¼ ðω1 ; …; ωp Þ is
called the frequency base. Since the solution x(t) is quasi-periodic, the basic frequencies wj must be incommensurate
P
(rationally independent), that is, for integers kj the equation 〈k; ω〉≔ pj¼ 1 kj ωj ¼ 0 holds if and only if all kj ¼ 0 for j ¼ 1; …; p.
For p ¼2 this means that the ratio ω1 =ω2 is irrational. In this paper, we use the term “quasi-periodic response” for a quasi-
periodic response with a two-dimensional frequency base.
Quasi-periodic responses naturally result from quasi-periodically forced nonlinear systems with a two-dimensional
frequency base. For example, multiple-rotor systems in aeroengines are subject to multi-frequency excitation due to several
coexisting unbalances [1]. For such nonlinear systems, the response possesses the same fundamental frequency base as the
external excitation, whose frequency base is determined. The dynamic response can be thus predicted by the numerical
algorithm based on a multi-dimensional generalization of harmonic balance [2]. In this method, the quasi-periodic response
is represented by a multi-dimensional Fourier series. The problem of predicting a quasi-periodic response is converted into
the problem of seeking a solution of an algebraic system of equations. This solution provides the amplitudes associated with
the different frequency spectral peaks in the assumed Fourier series. However, a conventional multi-dimensional harmonic

n
Corresponding author.
E-mail addresses: biao.zhou@nuaa.edu.cn (B. Zhou), fabrice.thouverez@ec-lyon.fr (F. Thouverez), david.lenoir@ec-lyon.fr (D. Lenoir).

http://dx.doi.org/10.1016/j.ymssp.2015.04.022
0888-3270/& 2015 Elsevier Ltd. All rights reserved.
234 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

balancing requires all the basic frequencies as a prerequisite; on the other hand, the concept of hyper-time is used to extend
the harmonic balancing to quasi-periodicity; the associated computational cost increases significantly with the number of
harmonics.
Apart from the multi-dimensional harmonic balance method, there also exist a number of methods for constructing
quasi-periodic solutions [5]. In general, analytical methods such as perturbation methods, multiple scales method and
averaging method are limited in small-sized systems possessing certain types of nonlinearities. Poincaré maps play a key
role in global analyses; the construction of high-order Poincaré sections depends on the availability of a large amount of
data and on the closeness of the ratio of the incommensurate frequencies to a rational number; hence, an adequate
knowledge of the nonlinear system (special emphasis on the basic frequencies) is required.
In particular, quasi-periodic responses have also been observed in nonlinear systems undergoing single-frequency
excitation. A well-known case is that rotors with bearing clearances and cross-coupling stiffness coefficients may exhibit
quasi-periodic response under a mono-harmonic excitation [3,6,7]. A simplified Jeffcott rotor model is usually adopted to
gain insights into many important phenomena observed in real applications. The transition from no contact to contact is
described by non-smooth nonlinear restoring forces. Cross-coupling stiffness might lead to a new output frequency
component, which is responsible for the outcoming of quasi-periodic responses. Research efforts have been focused on
frequency domain approaches to obtain steady state responses and understand the transition between periodic and quasi-
periodic responses. Provided that the extra-unknown basic frequency can be determined by, e.g., applying FFT algorithm to
the time history of the response, the quasi-periodic whirling response can be obtained similarly by means of multi-
dimensional harmonic balance method [3] or the so-called quasi-periodic harmonic balance method [8] in conjunction with
pseudo arc length continuation. However, methodologies of this kind are initiated by time integration simulations. The
prerequisite is to properly approximate the unknown basic frequency with an acceptable computational cost, which remains
a challenge in the literature.
Quasi-periodic response has also been widely investigated in the studies of nonlinear energy sinks (NES) in recent years.
A SDOF linear oscillator weakly coupled to a local small mass through strong stiffness nonlinearity, which acting as
nonlinear energy sink composes the simplest NES system (see Section 3). It is reported that even in the case of single-
frequency excitation, quasi-periodic response regimes are demonstrated to typically exist together with periodic response
regimes [9]. The new-arising basic frequency due to nonlinear coupling is not determined. In addition to the time marching
method, only analytical approach has been proposed for the strongly modulated response description [4].
In this paper, a new numerical approach, called Variable-Coefficient Harmonic Balance Method (VCHBM), will be
developed to predict the aforementioned quasi-periodic response precisely. The unknown basic frequency in the quasi-
periodic response is not necessarily predetermined, which advantages this new VCHBM over the multi-dimensional
harmonic balance method or the so-called quasi-periodic harmonic balance method. This method derives from the classical
harmonic balancing and its essential idea is to treat the quasi-periodic response as a truncated trigonometric series with
time-periodic coefficients. Harmonic balancing procedure is performed with respect to each basic frequency involved in the
quasi-periodic response. These time-periodic coefficients can then be sought by solving an algebraic system of equations
resulting from the previous harmonic balancing. As an effective numerical method currently well-developed and optimized
even for large scale nonlinear systems, the harmonic balance method is employed for seeking these variable Fourier
coefficients. To this end, the conventional harmonic balance method is first presented in Section 2, which serves as the
framework of harmonic balancing; then the principle of VCHBM is outlined and practical aspects concerning this numerical
algorithm are discussed; numerical examples are given in Section 3 in order to validate this newly developed method;
conclusions are drawn in the end.

2. Variable-coefficient harmonic balance method for quasi-periodic response

In this section, a variable-coefficient harmonic balance method is developed to characterize the quasi-response of
nonlinear systems. Let us consider a non-autonomous nonlinear dynamic system described by the following second order
differential equation:
Mx€ þ Dx_ þ Kx þ f nl ðt; x; xÞ
_ ¼ fðtÞ ð1Þ
_ represents nonlinear force
where M, D and K are mass, damping and stiffness matrix of size n  n, respectively; f nl ðt; x; xÞ
and fðtÞ the external periodic excitation with a single basic frequency ω. Since the Harmonic Balance Method (HBM) is
employed as a principal tool in this paper, it will be introduced in detail. Continuation technique is also covered in order to
follow the branch of solutions by HBM.

2.1. Harmonic balance method and continuation technique

In majority of cases, periodic response of nonlinear systems is observed. n-dimensional periodic solutions x(t) to Eq. (1)
of a unique fundamental frequency ω can be expressed by a truncated Fourier series:
X
Nh
xðtÞ ¼ X0 þ Xck cos ðkωtÞ þ Xsk sin ðkωtÞ ð2Þ
k¼1
 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244 235

where Nh is the number of retained harmonics. A ð2N h þ1Þn  1 vector X composed of Fourier coefficients of displacement is
defined as

X ¼ ½X0 ; Xc1 ; Xs1 ; …; Xck ; Xsk ; …T ð3Þ

If the trigonometric functions are regrouped as
TðωtÞ ¼ ½I; cos ðωtÞI; sin ðωtÞI; …; cos ðkωtÞI; sin ðkωtÞI; … ð4Þ
where I is the n  n identity matrix. Eq. (11) can be rewritten in a compact form:
xðtÞ ¼ TðωtÞX ð5Þ
A frequential derivative operator is introduced as
 
0 I
∇ ¼ diagð0nn ; ∇1 ; …; ∇Nh Þ with ∇k ¼ k ð6Þ
I 0

so that we can express the velocity and acceleration as

_ ¼ ωTðωtÞ∇X
xðtÞ ð7Þ

€ ¼ ω2 TðωtÞ∇2 X
xðtÞ ð8Þ
For the external force f and nonlinear term f nl , we have similar expressions following Eq. (5):

fðtÞ ¼ TðωtÞ  ½F0 ; Fc1 ; Fs1 ; …; Fck ; Fsk ; …T ¼ TðωtÞF ð9Þ

_ ¼ TðωtÞ  ½F0nl ; Fc1

f nl ðt; x; xÞ nl ; Fnl ; …; Fnl ; Fnl ; … ¼ TðωtÞFnl ðXÞ
s1 ck sk T
ð10Þ

The Galerkin procedure is then applied to the problem described by Eq. (1), i.e., projecting the equation on trigonometric
functions [10,11]:
Z 2π =ω
ðMx€ þ Dx_ þKx þf nl ðt; x; xÞ
_  fðtÞÞ dt ¼ 0
0
Z 2π =ω
_  fðtÞÞ cos ðkωtÞ dt ¼ 0
ðMx€ þ Dx_ þKx þf nl ðt; x; xÞ
0
Z 2π =ω
_  fðtÞÞ sin ðkωtÞ dt ¼ 0
ðMx€ þ Dx_ þKx þf nl ðt; x; xÞ
0

Substituting Eqs. (5)–(10) into the above equations yields a set of nonlinear algebraic equations linking Fourier
coefficients:

ðω2 NM ∇2 þ ωND ∇ þ NK ÞX þ Fnl ðXÞ  F ¼ 0 ð11Þ

or in a compact form:
Gðω; XÞ ¼ PðωÞX þ Fnl ðXÞ  F ¼ 0 ð12Þ
where the dynamic stiffness matrix PðωÞ ¼ ω NM ∇ þ ωND ∇ þNK . Note that NM is a block-diagonal matrix built as
2 2

NM ¼ diagðM; M; …Þ. ND and NK are constructed in the same way, respectively.

Roots of Eq. (12) are found by a quasi-Newton algorithm. The problem of searching for periodic solutions to the nonlinear
dynamic system equation (1) is therefore converted to an equivalent root-finding problem. The solving process needs to
evaluate Fnl ðXÞ at each iterative step. The alternating frequency/time-domain (AFT) technique [12] on the basis of Fast
Fourier Transforms (FFT) is outlined below for the determination of nonlinear Fourier coefficients Fnl ðXÞ:
TðωtÞX FFT
X ⟶ xðtÞ; _
xðtÞ⟶f _
nl ðt; x; xÞ⟶ Fnl ðXÞ ð13Þ
If the system behavior over a frequency band is required, one can use continuation technique [10] to find other solutions
and plot the response curve when ω is varied consecutively. Note that G is a nonlinear function taking its values in a space of
dimension ð2Nh þ1Þn þ 1. Continuation technique consists in numerically path-following a series of points ðωj ; Xj Þ from an
initial point ðω0 ; X0 Þ, so that the convergence criterion Gðωj ; Xj Þ r ξ is satisfied, with ξ being a prescribed accuracy [13]. In an
arc length continuation, solutions are parameterized by the arc length s, i.e., both ω and X are considered to be a function of
s so that ðωj ; Xj Þ ¼ ðωðsj Þ; Xðsj ÞÞ. Using the notation GX ¼ ∂G=∂X and Gω ¼ ∂G=∂ω, one obtains from Eq. (12):
" 0#
dX dω X
GX þ Gω ¼ ½GX Gω  ¼0 ð14Þ
ds ds ω0
with GX ¼ PðωÞ þ∂Fnl =∂X and Gω ¼ ð∂P=∂ωÞX.
236 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

An additional normalization condition is usually put forward so that the tangent vector ½X0 ; ω0 T has unit length:
‖dX‖2 ‖dω‖2
þ ¼1 ð15Þ
‖ds‖2 ‖ds‖2
Eqs. (14) and (15) form an implicit system of ð2Nh þ1Þn þ 1 differential equations for the ð2N h þ 1Þn þ1 unknowns dX and
dω. The solution ðωðsj Þ; Xðsj ÞÞ could then be fixed at a certain discretized arc length distance ds ¼ sj  sj  1 .
The combination of Eqs. (12) and (15) allows the solver to move along the arc-length s of a solution branch. A predictor–
corrector and adaptive step control scheme [13] has been adopted to enhance and speed up the path-following process in
order to search periodic solutions to Eq. (1) along a range of values for ω. Stability of periodic solutions is determined by the
Floquet theory [5] in this research.
In the literature, the harmonic balance method has been shown computationally efficient for simple nonlinear problems
as well as complex industrial nonlinear systems [14,15].

2.2. Variable-coefficient harmonic balance method

For all linear systems subject to mono-harmonic excitation, their responses will also be mono-harmonic. But this
consideration cannot be directly applied to nonlinear systems. Sometimes it happens that nonlinear response is quasi-
periodic or even chaotic. For the particular cases aforementioned in Introduction (also see Section 3), it is revealed that
periodic solutions on the response curve may evolve into quasi-periodic solutions. In other words, there emerges a new
unknown frequency component in the response. It renders the calculation of quasi-periodic solutions to Eq. (1) rather
difficult. A numerical approach, termed Variable-Coefficient Harmonic Balance Method (VCHBM), is proposed accordingly to
predict the quasi-periodic response precisely.
Variable-coefficient harmonic balance method stems from ideas in the method of multiple scales and harmonic balance
initially reported in 1995 [16]. In recent years, similar ideas can also be found as Fourier method by Schilder [17]. In the
Fourier method, a generalization of the averaging method is proposed to have an approximation of quasi-periodic invariant
tori where quasi-periodic solutions take place in phase space.
In essence, the VCHBM aims to reformulate the non-autonomous system giving rise to quasi-periodic response into an
autonomous system experiencing periodic response. First of all, it is necessary to briefly introduce the cornerstone of this
method. One can expand a quasi-periodic response xðtÞ into a Fourier series and collect the terms in ω2 as follows:
0 1
X1 X
1 X
1
iðk1 ω1 þ k2 ω2 Þt
xðtÞ ¼ ck1 ;k2 e ¼ @ ck1 ;k2 e 2 2 Aeik1 ω1 t
ik ω t

k1 ;k2 ¼  1 k1 ¼  1 k2 ¼  1

X
1
¼ uk1 ðω2 tÞeik1 ω1 t ð16Þ
k1 ¼  1

This manipulation shows that a 2-dimensional quasi-periodic solution can be approximated by such a generalized (or
multi-dimensional) Fourier series where the coefficients uk1 ðω2 tÞ are 2π periodic functions. For quasi-periodic solutions of
the nonautonomous system Eq. (1), an approximation xðω1 t; ω2 tÞ could be written equivalently by truncated trigonometric
functions:
X
Nh
xðω1 t; ω2 tÞ ¼ X0 ðω2 tÞ þ Xck1 ðω2 tÞ cos ðk1 ω1 tÞ þ Xsk1 ðω2 tÞ sin ðk1 ω1 tÞ
k1 ¼ 1

¼ ½I; cos ðω1 tÞI; sin ðω1 tÞI; …; cos ðk1 ω1 tÞI; sin ðk1 ω1 tÞI ,... 
½X0 ðω2 tÞ; Xc1 ðω2 tÞ; Xs1 ðω2 tÞ; …; Xck1 ðω2 tÞ; Xsk1 ðω2 tÞ; …T
¼ T1 ðω1 tÞXðω2 tÞ ð17Þ

where the external forcing frequency is denoted by ω1. The Fourier coefficient vector Xðω2 tÞ with an unknown frequency
component ω2 permits amplitudes of the assumed Fourier modes to vary with time. It principally characterizes the VCHBM
from the classical harmonic balance method in Eq. (5), where the Fourier coefficient vector X is constant. In the following
text, the VCHBM will be outlined as a two-step scheme.

2.2.1. Step 1: separation of frequency components

The first step of VCHBM is to separate the time-varying amplitudes Xðω2 tÞ from the nonlinear system in Eq. (1) to
eliminate the explicit occurrence of ω1 t. This is achieved by applying the harmonic balancing in terms of ω1 to the system in
Eq. (1). Starting from Eq. (17), the velocity and acceleration could be expressed as
_ ¼ ω1 T1 ∇1 X þ T1 X0
xðtÞ ð18Þ

€ ¼ T1 ðX00 þ 2ω1 ∇1 X0 þ ω21 ∇21 XÞ

xðtÞ ð19Þ
0
where X ¼ ∂X=∂t; X″ ¼ ∂ X=∂t . Subscripts of T1 and ∇1 indicate that they are defined in terms of
2 2
ω1 (see Eqs. (6)–(8)).
 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244 237

The Galerkin projection onto the subspace spanned by the trigonometric functions in T1 ðω1 tÞ is then performed for the
nonlinear system in Eq. (1):
NM ðX″ þ2ω1 ∇1 X0 þ ω21 ∇21 XÞ þ ND ðω1 ∇1 X þX0 Þ þ NK X þ Fnl ðXÞ  F ¼ 0 ð20Þ
_ ¼ T1 Fnl ðXÞ and fðtÞ ¼ T1 F.
where f nl ðt; x; xÞ
Physical meaning of the outcoming of harmonic balancing in terms of ω1 would become clear if the above equation is
rearranged as below:
NM X″ þ ð2ω1 NM ∇1 þ ND ÞX0 þðω21 NM ∇21 þ ω1 ND ∇1 þ NK ÞX þ Fnl ðXÞ  F ¼ 0 ð21Þ
It is an autonomous second order differential equation of the time-varying coefficient vector Xðω2 tÞ of size nð2N h þ1Þ in a
succinct form:
Ma X″ þDa X0 þ Ka X þ Fnl ðXÞ  F ¼ 0 ð22Þ
the subscript ðaÞ indicates that the matrices or vector correspond to the autonomous system.

2.2.2. Step 2: solving for periodic solution Xðω2 tÞ

The second step of VCHBM is to fix a periodic solution X to the autonomous system of Eq. (22) by standard algorithm.
Once again, the classical HBM, along with a suitable phase condition for such an autonomous system, are adopted in this
research. One can express the desired periodic solution by XðtÞ ¼ T2 ðω2 tÞZ, where Z is a Fourier coefficient vector of size
nð2Nh þ 1Þð2Nh2 þ 1Þ to be determined. Harmonic balancing procedure in terms of ω2 is then applied to Eq. (22) with Nh2
harmonics retained, which leads to
Gðω1 ; ω2 ; ZÞ ¼ Pðω1 ; ω2 ÞZ þ Fa;nl ðZÞ ¼ 0 ð23Þ
where Pðω1 ; ω2 Þ ¼ ω 2 NMa ∇2 þ
2 2
ω2 NDa ∇2 þ NKa and Fnl ðXÞ  F ¼ T2 ðω2 tÞFa;nl ðZÞ. The dynamic stiffness matrix Pðω1 ; ω2 Þ is
constructed in an analogous manner as in Eq. (12).
Notice that Eq. (23) does not define a unique periodic solution due to arbitrary phase shift. A well known integral phase
condition [5] can be added to this underdetermined equation (ω2 is an unknown parameter) in order to obtain a unique
periodic solution with a fixed phase shift:
Z 2π =ω2
hðω2 ; XÞ ¼ ^_ ðtÞ dt ¼ 0
XðtÞT X ð24Þ
0

Physical meaning of the integral phase condition is to minimize changes in a periodic solution profile [13]. In Eq. (24),
^
XðtÞ is a known nearby solution, usually the initial guess at the starting point, or consecutive points at the previous
^_ ðtÞ ¼ ω2 T2 ∇2 Z^ into this phase condition and
parameter value in a continuation. Substituting the relations XðtÞ ¼ T2 Z and X
further simplifications yield

hðω2 ; ZÞ ¼ ω2 ZT ∇2 Z^ ¼ 0 ð25Þ
In brief, periodic solution of the autonomous system equation (22) is fixed by solving the following set of algebraic
equations for nð2N h þ1Þð2Nh2 þ 1Þ þ 1 unknowns ðω2 ; ZÞ:
(
hðω2 ; ZÞ ¼ ω2 ZT ∇2 Z^ ¼ 0
Hðω1 ; ω2 ; ZÞ ¼ ð26Þ
Gðω1 ; ω2 ; ZÞ ¼ Pðω1 ; ω2 ÞZ þ Fa;nl ðZÞ ¼ 0

In addition, when solving the above equations, some practical aspects concerning this numerical algorithm have to be
taken into consideration.

2.2.3. Treatment of nonlinear term Fa;nl ðZÞ

It is noted that in Eq. (23) the nonlinear term Fa;nl ðZÞ is related to the time-varying Fourier coefficient vector Fnl ½XðtÞ of
_ is of a simple form (e.g., a cubic nonlinearity),
the nonlinear force f nl . In classical HBM, if the nonlinear function f nl ðt; x; xÞ
analytic expression of Fnl ½XðtÞ with respect to XðtÞ can be readily obtained for small-valued Nh. However, when a lot of
harmonics are included in the analysis, the resulting analytic expression may be long and complex. Consequently the
derivation of Fnl ½XðtÞ according to the classical HBM becomes cumbersome or even impractical in case that no analytical
expression can be written between Fnl ½XðtÞ and XðtÞ.
We hereby propose a variant of the alternating frequency/time domain (AFT) technique [12] specifically developed for
the treatment of nonlinear term Fa;nl ðZÞ. This procedure takes advantage of 2-Dimensional Fast Fourier Transform (2D FFT)
and the iterative step is sketched below:
T 1 T2 Z
Z ⟶ xðω1 t; ω2 tÞ; _ ω1 t; ω2 tÞ⟶f nl ðω1 t; ω2 t; x; xÞ
xð _
2D FFT FFT
⟶ f k1 ;k2 ⟶Fnl ½XðtÞ⟶Fa;nl ðZÞ ð27Þ

At each iterative step, an evaluation of the approximate temporal terms xðω1 t; ω2 tÞ and xð _ ω1 t; ω2 tÞ is carried out from an
initial value Z; it also allows to evaluate temporarily the nonlinear term f nl ðω1 t; ω2 t; x; xÞ.
_ 2D FFT is then performed to obtain
238 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

the Fourier coefficient in complex form f k1 ;k2 , which stands for the contribution of the frequency component k1 ω1 þk2 ω2 in
the time series of f nl ðω1 t; ω2 t; x; xÞ.
_ The time-varying Fourier coefficient vector Fnl ½XðtÞ is subsequently constructed. The key
point is that according to Eq. (16), the time-varying complex Fourier coefficient uk1 for the k1 ω1 harmonic term is
approximated as
Nh2
X
uk1 ðω2 tÞ ¼ f k1 ;k2 eik2 ω2 t ð28Þ
k2 ¼  Nh2

f nl ðtÞ is then expressed in terms of exponentials as

X
Nh
f nl ðtÞ ¼ uk1 eik1 ω1 t
k1 ¼  Nh

X
Nh
¼ u0 þ ðuk1 eik1 ω1 t þ u  k1 e  ik1 ω1 t Þ ð29Þ
k1 ¼ 1

Considering the property of Fourier series: u  k1 ¼ unk1 (the superscript n denotes complex conjugation), one can
manipulate the above expression in the following way:
X
Nh
f nl ðtÞ ¼ u0 þ ðuk1 eik1 ω1 t þunk1 e  ik1 ω1 t Þ
k1 ¼ 1

X
Nh
¼ u0 þ ½Reðuk1 Þðeik1 ω1 t þ e  ik1 ω1 t Þ
k1 ¼ 1

þiImðuk1 Þðeik1 ω1 t e  ik1 ω1 t Þ

X
Nh
¼ u0 þ ½2 Reðuk1 Þ cos ðk1 ω1 tÞ 2 Imðuk1 Þ sin ðk1 ω1 tÞ ð30Þ
k1 ¼ 1

Hence, the time-varying vector Fnl ½XðtÞ composed of real trigonometric Fourier coefficients reads as
0 1
Nh2
X
Fnl ðtÞ ¼ Reðu0 Þ ¼ Re@
0 ik ω
f 0;k2 e 2 2 A
t
ð31Þ
k2 ¼  N h2

N h2
X
Fcnlk1 ðtÞ ¼ 2 Reðuk1 Þ ¼ 2 Reðf k1 ;k2 Þ cos ðk2 ω2 tÞ
k2 ¼  N h2

 Imðf k1 ;k2 Þ sin ðk2 ω2 tÞ

Nh2
X
Fsnlk1 ðtÞ ¼ 2 Imðuk1 Þ ¼  2 Imðf k1 ;k2 Þ cos ðk2 ω2 tÞ
k2 ¼  Nh2

þ Reðf k1 ;k2 Þ sin ðk2 ω2 tÞ

The last step is to determine the Fourier coefficient vector Fa;nl ðZÞ in terms of ω2 from the approximation Fnl ½XðtÞ in the
time domain.
As a result, an explicit evaluation of Fnl ½XðtÞ is avoided and the VCHBM could be applied with a large number of
harmonics. Moreover, nonlinearities of all kinds could be treated by this variant of the alternating frequency/time domain
technique.

2.2.4. Initial condition

Solutions to the autonomous system in Eq. (22) are related to solutions to the non-autonomous system in Eq. (1) in the
following aspect: the autonomous system in Eq. (22) has a 2π periodic solution XðtÞ if and only if T1 XðtÞ is a quasi-periodic
solution to the non-autonomous system in Eq. (1). In reality, it is found that there coexist both a static solution ω2 ¼ 0 and a
dynamic solution ω2 a 0 to the autonomous system of Eq. (22). Normally, when solving Eq. (26) the static solution is always
achieved without any special treatment of the initial condition. This situation should be avoided for pursuing periodic
solutions. Hence, a sophisticated initial condition based on nonlinear stability analysis at the static equilibrium [18] is briefly
outlined.
First, nonlinear equation corresponding to the static equilibrium solution is described as
Ka Xs þ Fnl ðXs Þ  F ¼ 0 ð32Þ
 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244 239

Unbalance
U
Flexible, Massless Disk
Shaft
Gap
Flexible ks m,e kb
Bearing

δ
r y
kb ks
x

kb

Fig. 1. A Jeffcott rotor model with a bearing clearance.

Note that the static equilibrium Xs (a constant vector) is achieved when X0s ¼ 0 and X″s ¼ 0. The system equation (22) is then
linearized around the static equilibrium by the perturbation technique. Denoting X ¼ Xs þ ΔX, a linearized approximation of
system behavior in the neighborhood of the equilibrium is given by
Ma ΔX″ þ Da ΔX0 þ ½Ka þ JX Fnl ðXs ÞΔX ¼ 0 ð33Þ
where JX Fnl ðXs Þ is a Jacobian matrix of the nonlinear term Fnl ðXÞ evaluated at Xs .
The expression above is then written in the state-space form and complex eigenvalues θ ¼ a þ iω of the matrix
" #
0 I
A¼ 1 1 ð34Þ
 Ma ½Ka þ JX Fnl ðXs Þ Ma Da

are derived. Eigenvalues θ provide information about the local stability of the nonlinear system in case of small
perturbation. A positive real part of the eigenvalue equivalent to a negative damping leads to an unstable mode.
Consequently, the static equilibrium may lose stability and evolve into a periodic solution.
As explained by Sinou [18], the evolution of the dynamical solution near the static equilibrium can be approximated by
the unstable mode as below

Xðt; p; λÞ ¼ qðΨeθt þ Ψ eθ t Þ ð35Þ

where Ψ defines the nonlinear unstable mode and Ψ denotes its conjugate; q is an arbitrary coefficient. In this research, the
expression equation (35) is set as initial condition for solving equation (26); the unstable mode frequency is chosen as an
initial guess of the unknown frequency ω2, as well. The trivial static solution is therefore evitable during searching for
periodic solutions.

2.2.5. VCHBM in combination with arc length continuation

The quasi-periodic response branch could also be traced by the arc length continuation technique when the excitation
frequency ω1 is varied sequentially. Recall that the problem of searching for quasi-periodic solutions is reduced to solving a
collection of algebraic equations defined in Eq. (26). Denote the nð2N h þ1Þð2Nh2 þ1Þ þ 1 unknowns as Y ¼ ðω2 ; ZÞ. As before,
once an initial point ðω01 ; Y 0 Þ is found that satisfies Hðω01 ; Y0 Þ ¼ 0, we aim to consecutively locate a series of points ðωj1 ; Y j Þ
that minimize the residual H. Similar to Eqs. (14), (15), the incremental vector ðdω1 ; dYÞ along the solution branch in a
nð2Nh þ 1Þð2Nh2 þ 1Þ þ 2 dimensional space is determined as follows:
dω1 dY
Hω1 þ HY ¼0 ð36Þ
ds ds

‖dω1 ‖2 J dY J 2
þ ¼1 ð37Þ
‖ds‖2 ‖ds‖2
with the following derivatives calculated by numerical differentiations:
" # " #
∂h=∂ω1 0
Hω1 ¼ ¼ ð38Þ
∂G=∂ω1 ð∂P=∂ω1 ÞZ

" # " #
∂h=∂ω2 ∂h=∂Z ZT ∇2 Z^ ^ T
ω2 ð∇2 ZÞ
HY ¼ ¼ ð39Þ
∂G=∂ω2 ∂G=∂Z ð∂P=∂ω2 ÞZ P þ∂Fa;nl =∂Z

Conventional arc length continuation procedure can then be carried out to generate a continuum of points ðωj1 ; Y j Þ.
240 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

4 10
TM TM
3.5 VCHBM HBM

3 5

2.5

2 0

1.5

1 −5

0.5

0 −10
0.2 0.4 0.6 0.8 1 1.2 1.4 1900 1950 2000 2050 2100 2150 2200
Frequency nondimensional time
Fig. 2. Numerical validation of quasi-periodic response in the Jeffcott rotor model: α ¼ 1, ζ ¼ 0:2, Ω ¼ 4:8, γ ¼ 0:4, ϕ ¼ 1, δn ¼ 10; (a) X 1 in the frequency
domain; (b) x 1 ðtÞ in the time domain; TM: solution by time marching; VCHBM: solution by the variable coefficient harmonic balance method.

Above all, once a periodic solution Xðω2 tÞ is sought, the quasi-periodic response is computed by xðtÞ ¼ T1 ðω1 tÞXðω2 tÞ.
Stability analysis of quasi-periodic solutions is not currently included. At this point, the numerical algorithm for the VCHBM
is well developed and could be used for the prediction of quasi-periodic response in numerical examples presented in the
next section.

3. Numerical example

In this section, the VCHBM will be applied to rotor dynamics and NES systems in order to demonstrate its effectiveness in
predicting quasi-periodic response with an unknown basic frequency.

3.1. VCHBM applied to rotor dynamics

Consider a horizontal Jeffcott rotor model with a bearing clearance and cross-coupling forces (see Fig. 1). The governing
equation of motion is described as [3]
0 1
B δ C
mx€ 1 þ cx_ 1 þ ks x1 þ Q s x2 þ ϕkb x1 @1  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiA ¼ meω2 cos ðωt Þ ð40Þ
x1 þ x22
2

0 1
B δ C
mx€ 2 þ cx_ 2 þ ks x2  Q s x1 þ ϕkb x2 @1  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiA ¼ meω2 sin ðωt Þ  mg ð41Þ
x21 þ x22

where m is the disk mass, c is the damping coefficient, ks is the massless shaft stiffness, Qs is the cross coupling stiffness
coefficient, δ is the bearing clearance, kb is the bearing stiffness, e is the mass eccentricity, ω is the rotational speed. ϕ takes
the form of
8 qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
< 0; δ Z x21 þ x22 no contact
ϕ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð42Þ
>
: 1; δ o x21 þ x22 contact occurs

A series of nondimensional quantities are introduced as follows:

pffiffiffiffiffi pffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffi
x 1 ¼ x1 =e; x 2 ¼ x2 =e; K ¼ 4ks kb =ð ks þ kb Þ2 ; ωn ¼ K=m; Ω ¼ ω=ωn
ζ ¼ c=2 mωn ; γ ¼ Q s =K; α ¼ kb =ks ; ψ ¼ g=eω2n ; t ¼ ωt ð43Þ

By means of these listed quantities above, Eqs. (40)–(41) are then nondimensionalized in the following form:
pffiffiffiffi
2ζ ð1 þ αÞ2 γ  
x€ 1 þ x_ 1 þ x 1 þ 2 x 2 þ T t ¼ cos t ð44Þ
Ω 4αΩ
2
Ω
pffiffiffiffi
2ζ ð1 þ αÞ2 γ  
x€ 2 þ x_ 2 þ x 2  2 x 1 þ N t ¼ sin t  ψ =Ω
2
ð45Þ
Ω 4αΩ
2
Ω
 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244 241

Table 1
System parameters.

ε λ knl A

0.1 0.4 5 0.3

Differentiation in Eqs. (44)–(45) is with respect to nondimensional time t . The nonlinear terms Tðt Þ and Nðt Þ are
expressed by
pffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Tðt Þ ¼ Φx 1 ð1=Ω Þð1 þ αÞ2 =4  ð1  δ = x 21 þ x 22 Þ
2 n
ð46Þ

pffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Nðt Þ ¼ Φx 2 ð1=Ω Þð1 þ αÞ2 =4  ð1  δ = x 21 þx 22 Þ
2 n
ð47Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where Φ is unity if δ o x 1 þ x 2 and otherwise is zero.
n 2 2

Systematic analysis reveals the occurrence of a quasi-periodic whirling response in the Jeffcott model due to the presence
cross-coupling stiffness [3,6,7]. This phenomenon is observed in our numerical simulation results. As shown in Fig. 2a, this
quasi-periodic response is characterized by two incommensurate frequencies: one is the excitation frequency ω1 ¼ 1; the
other is an extra-unknown basic frequency ω2 to be determined. The two frequencies ω1 and ω2 are said to be
incommensurate because the ratio ω1 =ω2 is an irrational number.
The variable coefficient harmonic balance method is then applied to Eqs. (44)–(45) for the quasi-periodic response. The
retained harmonics in the VCHBM are set as Nh ¼5 in terms of the excitation frequency ω1 and N h2 ¼ 6 in terms of the new
frequency component ω2, respectively. A preliminary convergence study with the number of retained harmonics is carried
out. It reveals that a good approximation can be obtained by the chosen values of Nh and Nh2 . The unknown basic frequency
is identified by VCHBM as ω2 ¼ 0:208. For the VCHBM, it is noted that the size of Z increases rapidly (up to 286) due to more
harmonics involved. Since the periodic Fourier coefficient vector is eventually sought by the conventional HBM, a harmonic
selection procedure is naturally desired to lower down the computational cost. It aims to a dynamic management of the
predominant harmonics really contributing to the quasi-periodic solution in the application of HBM so that a balance is
achieved between the solution accuracy and the total amount of computational work. Such harmonic selection technique is
currently attracting more and more attention [8,15]. However, it is still under investigation for the VCHBM and not involved
in this paper. Despite the increasing size of Z, the overall computation time for the present academic examples remains an
acceptable level. The number of iterative step sketched in Eq. (27) accounts to 13 in this example.
In Fig. 2b, it can be seen that excellent agreement of the solution by VCHBM and the time-marching results is achieved. It
indicates that the VCHBM yields good accuracy in predicting the quasi-periodic whirling response. The main advantage of
the VCHBM is that it is capable of predicting quasi-periodic solutions including an unknown basic frequency. Accordingly, it
allows one to directly obtain the steady state response in the frequency domain. In contrast, a conventional multi-
dimensional generalization of harmonic balancing requires all the basic frequencies as a prerequisite. Numerical integration
in the time domain is usually performed to give an overview of the response spectrum in order to initialize the multi-
dimensional harmonic balancing [3,8]. However, this procedure is rather time-consuming for large scale systems.

3.2. VCHBM applied to NES systems

Nonlinear energy sink systems have been widely investigated in recent decades [19]. In-depth studies have reported that
quasi-periodic response regimes characterize the nonlinear frequency response curve of NES systems under a single
frequency excitation [20]. The simplest system capable of performing as NES is constructed as follows [21]:

x€ 1 þ ελðx_ 1  x_ 2 Þ þ x1 þ εknl ðx1 x2 Þ3 ¼ εA cos ðωtÞ ð48Þ

εx€ 2 þ ελðx_ 2  x_ 1 Þ þ εknl ðx2  x1 Þ3 ¼ 0 ð49Þ

This system consists of a grounded SDOF linear oscillator, which acts as the primary system with mass m1 ¼ 1, coupled to
an ungrounded attachment m2 ¼ ε through a pure cubic stiffness in parallel to a viscous damper. The equation of motion for
this integrated system clearly characterizes NES by

 strong mass asymmetry, ε 5 1;

 essentially nonlinear coupling between the primary structure and NES.

The system parameter set listed in Table 1 is chosen in this numerical example.
It has been demonstrated that the addition of a relatively small and spatially localized attachment (NES) significantly
alters the global dynamics of the resulting integrated system. Particularly, early analytical studies have revealed that the NES
is able to react efficiently on the amplitude characteristics of the external forcing in a wide range of frequencies [4]. It is
consequently desirable to explore the NES as the strongly nonlinear vibration absorber for periodic external loadings.
242 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

0.5

Im(ρ)
0

−0.5

−1
−1 −0.5 0 0.5 1
Re(ρ)
Fig. 3. Scenarios depicting of Floquet multipliers ( ) leaving the unit circle in the complex plane around ω ¼ 0:948 (Neimark–Sacker bifurcation).

0.3 0.7

0.6
0.25 NS
NS 0.5
Amplitude of x1

Amplitude of x2
0.2
0.4

0.3
0.15

0.2 NS NS

0.1
0.1

0.05 0
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2 0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2
Frequency (ω: rad/s) Frequency (ω: rad/s)

Fig. 4. Nonlinear frequency response: (a) Linear oscillator: x1; (b) NES: x2. ( : stable solution; : unstable solutions); NS: Neimark–Sacker
bifurcation point.

Nevertheless no frequency-domain approach for precise prediction of nonlinear response arising in such strongly nonlinear
systems has been reported.
In this section, the forced response of the coupled nonlinear system under single frequency external excitation is first
calculated by HBM and continuation technique with Nh ¼3 harmonics retained. Stability of periodic solutions manifests
itself through Floquet multipliers ρ. Depending on where the critical multiplier or pair of complex conjugate multipliers
crosses the unit circle in the complex plane, different types of bifurcation might occur. The Floquet multipliers of the
periodic solutions around ω ¼ 0:948 are depicted in Fig. 3. It is clearly seen that pairs of conjugate eigenvalues cross the unit
circle. It indicates that a bifurcation from a periodic solution to a torus (Neimark–Sacker bifurcation or generalized Hopf
bifurcation) comes out.
By monitoring Floquet multipliers ρ, two Neimark–Sacker bifurcation points are totally observed in the nonlinear
frequency response curves (see Fig. 4). The branch of stable periodic solutions that exists prior to the first bifurcation point
on the left continues as a branch of unstable periodic solutions after the bifurcation. At these bifurcation points, there
emerges a new unknown frequency component and the periodic response accordingly evolves into quasi-periodic
responses. It should be pointed out that the quasi-periodic response occurs due to the nonlinear coupling between
oscillation modes although the system is single-frequency excited. This result is in agreement with the literature, where
analytical approximation for boundaries of a strong and weak quasi-periodic modulation possibility is obtained by
Starosvetsky and Gendelman [21].
Quasi-periodic responses are then sought by the variable-coefficient harmonic balance method at the bifurcation point,
where periodic responses lose their stability. They are further followed by means of arc-length continuation when ω is
varied between the two NS bifurcation points. The retained harmonics Nh2 in the VCHBM is set as 5. In general, nonlinearity
in the term Fnl ½XðtÞ becomes stronger than that in the term f nl ðω1 t; ω2 t; x; xÞ.
_ Hence, more harmonics should be retained
(Nh2 Z Nh ) to have a good approximation of the system behavior. The size of Z is 154 for the given values of Nh and N h2 . The
number of iterative step sketched in Eq. (27) accounts to 5 for the calculation of an initial point, and reduces to 2–3 during
the arc length continuation. The complete nonlinear frequency response curves are depicted in Fig. 4 and numerical
validation of the periodic/quasi-periodic responses in the time domain is shown in Fig. 5. For quasi-periodic responses, the
identified frequency of modulation ω2 is also given below. It can be seen that the quasi-periodic solutions by VCHBM are of
good precision.
 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244 243

TM 0.3 TM
0.2 HBM VCHBM
0.15 0.2
0.1
0.1
0.05
x1(t)

x1(t)
0 0
−0.05
−0.1
−0.1
−0.15 −0.2

−0.2
−0.3
−0.25
0 5 10 15 20 25 0 50 100 150 200
time (t:s) time (t:s)

0.4 0.2
TM TM
0.3 VCHBM 0.15 HBM

0.2 0.1

0.1 0.05
x1(t)

x1(t)
0 0

−0.1 −0.05

−0.2 −0.1

−0.3 −0.15

−0.4 −0.2
0 50 100 150 200 0 5 10 15 20 25
time (t:s) time (t:s)

Fig. 5. Numerical validation of nonlinear frequency responses x1 ðtÞ. (a) Periodic response by HBM at ω ¼ 0:94. (b) Quasi-periodic response by VCHBM at
ω ¼ 0:99. (c) Quasi-periodic response by VCHBM at ω ¼ 1:03. (d) Periodic response by HBM at ω ¼ 1:07. TM: solution by time marching.

As introduced earlier, the quasi-periodic response regimes are very typical for the system of linear oscillator coupled with
NES under external harmonic forcing. It is therefore rewarding to see that the VCHBM works efficiently in predicting quasi-
periodic responses arising in NES systems. Since the NES is explored as a nonlinear vibration absorber, it is demonstrated
that such a quasi-periodic response regime usually provides more efficient performance in terms of vibration mitigation. A
prevailing advantage is that the NES can drastically reduce the vibration level while introducing no additional resonant
frequency [21,22].

4. Conclusion

This research focuses on the quasi-periodic responses arising from nonlinear systems under a single-frequency harmonic
forcing. Various coupling effects usually give rise to a new emerging basic frequency. A variable-coefficient harmonic
balance method is accordingly proposed for the prediction of quasi-periodic responses, in which there is an undetermined
basic frequency. This frequency-domain method exempts one from time-domain integrations, which is computationally
expensive for large scale nonlinear systems.
The essential idea of VCHBM is to treat the quasi-periodic response in a“cascade” of frequency base. Harmonic terms are
separated and balanced with respect to each frequency component. The process enables to construct a set of equations for
the time-varying coefficients. They are further evaluated by the conventional HBM, which is nowadays well developed and
optimized even for large scale nonlinear systems. The VCHBM greatly benefits from a specifically developed variant of the
alternating frequency/time domain technology based on 2D FFT so that it could be applied to various complex nonlinear
systems. The unknown basic frequency can therefore be identified as a by-product. Most importantly, no a priori knowledge
of the nonlinear system is required to initiate this methodology. This feature characterizes the VCHBM from the
conventional multi-dimensional generalization of harmonic balance, which deals with the quasi-periodic response globally
in the frequency base, and requires all the basic frequencies as a prerequisite.
Effectiveness of the VCHBM is tested in two cases where quasi-periodic response arises in rotor dynamics and NES
systems, respectively. Numerical simulations demonstrate that the variable-coefficient harmonic balance method is suitable
in predicting quasi-periodic responses in such nonlinear systems. Furthermore, the combination of VCHBM and continua-
tion technique could enhance the robustness and speed advantage of this frequency-domain methodology. However,
harmonic selection techniques that have potentiality to further lower down the computational cost is expected to be
244 B. Zhou et al. / Mechanical Systems and Signal Processing 64-65 (2015) 233–244

incorporated into the VCHBM. Finally, it should be reminded that the basic idea of the VCHBM is not limited to these tested
cases. Applications of the VCHBM to complex nonlinear systems (e.g., blade flutter instability problems under low engine
order excitation may experience quasi-periodic responses) are also desirable.

Acknowledgments

This project is supported by National Natural Science Foundation of China (Grant no. 51405460). This support is gratefully
acknowledged.

References

[1] R.R. Pusenjak, M.M. Oblak, Incremental harmonic balance method with multiple time variables for dynamical systems with cubic non-linearities, Int.
J. Numer. Methods Eng. 59 (2004) 255–292.
[2] M. Guskov, J.-J. Sinou, F. Thouverez, Multi-dimensional harmonic balance applied to rotor dynamics, Mech. Res. Commun. 35 (2008) 537–545.
[3] Y.B. Kim, S.T. Noah, Quasi-periodic response and stability analysis for a non-linear Jeffcott rotor, J. Sound Vib. 190 (1996) 239–253.
[4] O. Gendelman, Y. Starosvetsky, M. Feldman, Attractors of harmonically forced linear oscillator with attached nonlinear energy sink. I: description of
response regimes, Nonlinear Dyn. 51 (2008) 31–46.
[5] A.H. Nayfeh, B. Balachandran, Applied Nonlinear Dynamics: Analytical, Computational, and Experimental Methods, Wiley, New York, 1995.
[6] J. Jiang, H. Ulbrich, Stability analysis of sliding whirl in a nonlinear Jeffcott rotor with cross-coupling stiffness coefficients, Nonlinear Dyn. 24 (2001)
269–283.
[7] J.P. Chávez, M. Wiercigroch, Bifurcation analysis of periodic orbits of a non-smooth Jeffcott rotor model, Commun. Nonlinear Sci. Numer. Simul. 9
(2013) 2571–2580.
[8] L. Peletan, S. Baguet, M. Torkhani, G. Jacquet-Richardet, Quasi-periodic harmonic balance method for rubbing self-induced vibrations in rotor–stator
dynamics, Nonlinear Dyn. 78 (2014) 2501–2515.
[9] Y. Starosvetsky, O.V. Gendelman, Response regimes of linear oscillator coupled to nonlinear energy sink with harmonic forcing and frequency
detuning, J. Sound Vib. 315 (2008) 746–765.
[10] G. Groll, D. Ewins, The harmonic balance method with arc length continuation in rotor/stator contact problem, J. Sound Vib. 241 (2001) 223–233.
[11] V. Jaumouillé, J. Sinou, B. Petitjean, An adaptive harmonic balance method for predicting the nonlinear dynamic response of mechanical systems—
application to bolted structures, J. Sound Vib. 329 (2010) 4048–4067.
[12] T.M. Cameron, J.H. Griffin, An alternating frequency/time domain method for calculating the steady-state response of nonlinear dynamic systems,
J. Appl. Mech. 56 (1989) 149–154.
[13] R. Seydel, Practical Bifurcation and Stability Analysis: From Equilibrium to Chaos, Springer, New York, 1995.
[14] B. Cochelin, C. Vergez, A high order purely frequency-based harmonic balance formulation for continuation of periodic solutions, J. Sound Vib. 324
(2009) 243–262.
[15] A. Grolet, F. Thouverez, On a new harmonic selection technique for harmonic balance method, Mech. Syst. Signal Process. 30 (2012) 43–60.
[16] J.L. Summers, Variable-coefficient harmonic balance for periodically forced nonlinear oscillators, Nonlinear Dyn. 7 (1995) 11–35.
[17] F. Schilder, W. Vogt, S. Schreiber, H.M. Osinga, Fourier methods for quasi-periodic oscillations, Int. J. Numer. Methods Eng. 67 (2006) 629–671.
[18] J.J. Sinou, F. Thouverez, L. Jezequel, Stability analysis and non-linear behaviour of structural systems using the complex non-linear modal analysis
(CNLMA), Comput. Struct. 84 (2006) 1891–1905.
[19] A. Vakakis, O. Gendelman, Nonlinear Targeted Energy Transfer in Mechanical and Structural Systems, Springer, Netherlands, 2009.
[20] O.V. Gendelman, D.V. Gorlov, L.I. Manevitch, A.I. Musienko, Dynamics of coupled linear and essentially nonlinear oscillators with substantially
different masses, J. Sound Vib. 286 (2005) 1–19.
[21] Y. Starosvetsky, O. Gendelman, Attractors of harmonically forced linear oscillator with attached nonlinear energy sink. II: optimization of a nonlinear
vibration absorber, Nonlinear Dyn. 51 (2008) 47–57.
[22] E. Gourdon, N.A. Alexander, C.A. Taylor, C.H. Lamarque, S. Pernot, Nonlinear energy pumping under transient forcing with strongly nonlinear coupling:
theoretical and experimental results, J. Sound Vib. 300 (2007) 522–551.